HOAi Engineering
Recent blog posts.
-
Evaluating voice agent quality at scale
Sourish BiswasWe improved how our voice agent identifies callers, but a pass/fail metric couldn't see the difference. Here is how we graded thousands of calls with a second model, and what we found when we checked the grader against human review.
-
Improving caller identification for voice AI agents
Sourish BiswasWhen callers aren't recognized by phone number, our voice agent looks them up by name and address. That lookup succeeded 5 times out of 70. Here is how we fixed it.
-
Improving tail latency in parallel LLM calls
Varun SrinivasanParallel LLM calls made our document search faster on average and much slower at the tail. One straggler set the latency for the entire phase.
-
Debugging Postgres performance under Row-Level Security
Zhixuan LaiHow an inefficient Row-Level Security policy drove Postgres CPU to 80%; and why EXPLAIN failed to reveal it.
-
Why Postgres ignored our full-text search index
Varun SrinivasanOur full-text search index existed and Postgres maintained it correctly. The shape of our query made the planner ignore it.
-
Debugging stuck Node.js processes
Winona Schroeer-SmithOne of our critical apps kept failing health checks. Logs, metrics, and local repros came up empty. We dove into Node.js internals to find the root cause.
-
When the model overthinks
Zhixuan LaiSometimes reasoning models get stuck in loops that burn thousands of tokens and produce nothing. Here is how we fixed it.
-
A simple fix for LLM tail latency
Zhixuan LaiSend every request twice and take the faster response.
-
Just ask the model why
Zhixuan LaiWhen your agent does the wrong thing, ask it why. It usually knows.
-
How we evaluate AI agents across 1,000+ production scenarios
Pratik JainGeneric benchmarks can't tell you if your AI agent will make the right decisions in production. We built a replay-based evaluation framework that tests 1,000+ real scenarios to compare models and catch regressions before they reach 200,000 weekly workflow runs.